An Intelligent SIoT Trust Classification and Prediction Model Using Skip-GRU-Attention and Feed-Forward Networks

Trust evaluation in SIoT networks is increasingly complex, requiring advanced classification models. This study presents SAFE-SIoT, an Intelligent Trust Classification System (ITCS) that leverages Skip-GRU-Attention and Feed-Forward Neural Networks to model trust relationships non-linearly. The proposed framework enhances classification accuracy by incorporating centralization-based trust indicators, which improve differentiation among entities and optimize trust evaluation. SAFE-SIoT integrates Gated Recurrent Units (GRU) with skip connections to facilitate efficient gradient flow, an Attention Mechanism for prioritizing critical features in sequential data, and Extreme Learning Machines (ELM) for high-speed trust classification. By incorporating hierarchical attention mechanisms for deep feature extraction, the system enhances adaptability and robustness, making it highly suitable for secure and scalable SIoT environments. This paper introduces SAFE-SIoT, a novel trust classification model combining Skip-GRU, attention mechanisms, and ELM. It achieves 97.5% accuracy and demonstrates superior scalability and robustness compared to existing models. The results indicate that SAFE-SIoT significantly enhances trust evaluation and provides a scalable, high-performance solution for dynamic SIoT ecosystems. Future research will focus on further optimizing the model's adaptability to real-world applications and refining trust parameter integration to improve predictive capabilities. SAFE-SIoT introduces a novel hybrid deep learning framework tailored for trust evaluation in smart city Social Internet of Things (SIoT) environments. The architecture integrates Skip-GRU units to enhance temporal modeling by preserving long-range dependencies and mitigating vanishing gradients in sequential trust data. A Bi-Layered Attention (BLA) mechanism further refines feature prioritization by combining Position-Spatial Attention (PSA) and Channel Attention (CA), enabling the model to focus on both spatially salient regions and semantically rich channels. To ensure computational efficiency, the framework employs an Extreme Learning Machine (ELM) as a lightweight classifier, offering fast training and strong generalization—particularly beneficial for resource-constrained IoT nodes. SAFE-SIoT achieves 97.5% accuracy on benchmark SIoT trust datasets and demonstrates robust scalability across heterogeneous IoT deployments with minimal latency overhead. It consistently outperforms baseline models such as LSTM, CNN-GRU, and vanilla GRU, with precision, recall, and F1-scores of 96.8%, 97.2%, and 97.0%, respectively. Designed for secure and trustworthy communication, the model supports real-time trust evaluation in dynamic and decentralized SIoT topologies and is compatible with federated learning extensions to enable privacy-preserving deployment.